Re: [PATCH v2 5/5] arm64: dts: uniphier: Add NX1 SoC and boards support

[Kunihiko Hayashi <hayashi.kunihiko@socionext.com>]

On 2022/11/11 23:40, Krzysztof Kozlowski wrote:
> On 11/11/2022 09:48, Kunihiko Hayashi wrote:
>> Hi Krzysztof,
>>
>> On 2022/11/09 0:11, Krzysztof Kozlowski wrote:
>>> On 08/11/2022 15:30, Kunihiko Hayashi wrote:
>>>> Hi Krzysztof,
>>>>
>>>> On 2022/11/08 20:13, Krzysztof Kozlowski wrote:
>>>>> On 07/11/2022 11:34, Kunihiko Hayashi wrote:
>>>>>> Initial version of devicetree sources for NX1 SoC and boards.
>>>>>>
>>>>>> NX1 SoC belongs to the UniPhier armv8 architecture platform, and is
>>>>>> designed for IoT and AI/ML application fields.
>>>>>>
>>>>>
>>>>>> +
>>>>>> +		soc_glue: syscon@1f800000 {
>>>>>> +			compatible = "socionext,uniphier-nx1-soc-glue",
>>>>>> +				     "simple-mfd", "syscon";
>>>>>> +			reg = <0x1f800000 0x2000>;
>>>>>> +
>>>>>> +			pinctrl: pinctrl {
>>>>>> +				compatible = "socionext,uniphier-nx1-pinctrl";
>>>>>
>>>>> So instead of documenting the hardware precisily, you have one big bag
>>>>> for everything under simple-mfd. This is not how the SoC should be
>>>>> described in DTS.
>>>>
>>>> Sorry I don't understand. This is inherited from the previous
>>>> descriptions,
>>>> but is there some example to express DTS correctly about that?
>>>
>>> I think yes, although it actually depends what is this hardware.
>>> Generally speaking, do not use simple-mfd and syscon when these are not
>>> really simple devices. There are quite many in your DTS, which got my
>>> attention. Instead - have regular device with or without children.
>>>
>>> There is no real need to have this a simple-mfd with one children
>>> without any resources (no address space, no clocks, no interrupts,
>>> nothing).
>>>
>>> Why this syscon/mfd and pinctrl is not a regular, one device?
>>
>> The mfd/syscon.yaml says:
>>     System controller node represents a register region containing a set
>>     of miscellaneous registers.
>>
>> The "soc-glue" is exactly this, it contains various register functions
>> and might be referred to the drivers.
>>
>> For example in this NX1 dts, ethernet node points to "soc-glue" node.
>>
>>       eth: ethernet@15000000 {
>>           compatible = "socionext,uniphier-nx1-ave4";
>>           ...
>>           socionext,syscon-phy-mode = <&soc_glue 0>;
>>       };
>>
>> Since such register region is not often systematically designed,
>> it is tough to cut out as specific memory region for "pinctrl".
> 
> So your choice is instead use entire address space as pinctrl - as a
> child device without IO address space. That's also not a good solution.

This structure follow the existing UniPhier SoCs, so it is necessary
to re-think the structure of all SoCs, not just this NX1 SoC.

>>
>> And more, the existing pinctrl driver uses of_get_parent() and
>> syscon_node_to_regmap(), so this change breaks compatibility.
> 
> This is a new DTS, so what compatibility is broken? With old kernel?
> There was no compatibility with this Devicetree. Anyway using driver
> implementation as reason for specific hardware description (DTS) is also
> not correct.

In this same area, mode selector, clock selector, phy configuration etc,
exist together in a mixed manner. There is a kind of design issues.

 From this point of view, I should separate the new devicetree once
from this patch series, and it is necessary to consider syscon DT schema
and related drivers for existing SoCs.

>>>>>> +			};
>>>>>> +		};
>>>>>> +
>>>>>> +		soc-glue@1f900000 {
>>>>>> +			compatible = "simple-mfd";
>>>>>
>>>>> No, it is not allowed on its own. You need a specific compatible and
>>>>> bindings describing its children.
>>>>
>>>> I saw the definition of "simple-mfd" itself is only in mfd/mfd.txt.
>>>>
>>>> Currently there are only efuse devices as children, and this space means
>>>> nothing. I think it had better define the devices directly.
>>>
>>> You need to start describe the hardware. efuse is an efuse, not MFD.
>>> pinctrl is pinctrl not MFD + pinctrl.
>>
>> This region also has multiple functions, though, the efuse might be
>> cut out as specific region without "simple-mfd", unlike pinctrl.
> 
> simple-mfd itself does not mean region has multiple functions, but that
> children do not depend on anything from the parent device.
 >
> You over-use syscon and simple-mfd in multiple places. of course some of
> them will be reasonable, but now it does not.

It takes more time to review existing structures, especially patch 1/5.

Thank you,

---
Best Regards
Kunihiko Hayashi